Startup offers commercial service to remove guardrails from open-weight AI models

Abliteration.ai offers a commercial platform that hosts modified versions of open-weight AI models with their guardrails removed, including Z.ai’s GLM-5.3, which users can query via a web browser or API. The company describes its goal as enabling “offensive cyber, red-teaming, and agent testing work other models refuse to do,” arguing that defenders need the same tools as adversaries to improve security.

The startup was founded in late 2025 and officially incorporated in March 2026. By hosting pre-abliterated models, Abliteration.ai reduces the friction for users who would otherwise need to download and run the models themselves. In a hands-on test, a TechCrunch reporter created an account and queried an abliterated version of GLM-5.3, successfully obtaining outputs for a Python program to steal saved Chrome passwords and a protocol for culturing a dangerous human pathogen at home.

Abliteration.ai has not raised venture capital and is in talks to do so, according to co-founder Devon, who spoke on condition that his last name not be included. The company has several deals with major cloud providers and relies solely on customer revenue. It offers a moderation layer so customers can add their own guardrails, and has implemented minor guardrails such as blocking instructions for suicide, with plans to add more restrictions on violence.

Critics argue that making abliterated models widely accessible lowers barriers to misuse. Andrew Yoon, head of research at AI safety nonprofit CivAI, described abliteration as enabling models to “become a sociopath,” adding, “I do expect we will start to see edited, abliterated models being used for harm in the near future.” He suggested governments could require providers to run classifiers to detect and block harmful cyber and bioweapons activity, and that companies renting direct access to advanced GPUs should verify customer identities.

Abliteration.ai’s customers include early-stage red-teaming startups in the UK and Europe that help banks, airlines, and critical infrastructure firms improve their cybersecurity. Devon argued that abliterated models allow defenders to “move as fast as possible” and model bad actors to accelerate cybersecurity improvements.

Some in the cybersecurity industry question how consequential abliterated models are for defensive work. Ahmed Aly, CEO of agent red-teaming firm Fabraix, said his company relies more on fine-tuning open models than using abliterated ones, noting that abliteration can remove some of a model’s knowledge and capabilities. Alessio Lomuscio, chief technologist at Safe Intelligence, acknowledged a possible reduction in capabilities but still sees value in abliterated models for certain testing scenarios.